Intelligent driving method and system based on generative AI and quantum computing

Abstract

The embodiment of the invention provides an intelligent driving method and system based on generative AI and quantum computing. According to the method, on the basis of current driving data of a vehicle, a generative AI model is utilized to predict a multi-modal behavior track of a traffic participant in a future preset time period, and a track prediction result is obtained; based on the trajectory prediction result, the map traffic data and the current driving data, solving the road network quantum diagram by using a quantum optimization algorithm to obtain a path planning result; the path planning result is converted into a drive-by-wire chassis execution instruction, and the vehicle is controlled to complete the driving behavior; the generative AI model is obtained by training a diffusion model adopting a space-time attention mechanism based on a driving planning open source data set, an extreme scene data set and a traffic flow data set; and the extreme scene data set is obtained by performing dynamic scene enhancement based on the GAN, so that the accuracy and the real-time performance of path planning can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent driving, and in particular to an intelligent driving method and system based on generative AI and quantum computing. BACKGROUND

[0002] Intelligent driving is a technology that enables autonomous or semi-autonomous operation of vehicles through on-board sensors, algorithms, and control systems. Its core goal is to improve driving safety, efficiency, and comfort.

[0003] In current driving methods based on intelligent driving technology, vehicles use a fixed road information to model the current road environment of the vehicle using a deep learning model, and then use Dijkstra algorithm to process the environment model to obtain a driving planning path. The deep learning model mainly relies on fixed information, but in the driving process, there are often other vehicles or pedestrians and other traffic participants, which makes it difficult for traditional driving schemes to adapt to dynamic scenarios. Moreover, the Dijkstra algorithm has high computational complexity in high-dimensional road networks, making it difficult to meet the real-time requirements of path planning.

[0004] Therefore, there are problems of low accuracy and real-time performance in the traditional driving scheme. SUMMARY

[0005] The embodiments of the present application provide an intelligent driving method and system based on generative AI and quantum computing to solve the existing problems.

[0006] The embodiments of the present application provide an intelligent driving method based on generative AI and quantum computing, comprising: Based on the current driving data of the vehicle, a generative AI model is used to predict the multi-modal behavior trajectory of the traffic participants in a future preset time period to obtain a trajectory prediction result; Based on the trajectory prediction result, the obtained map traffic data and the current driving data, a quantum optimization algorithm is used to solve a pre-constructed road network quantum graph to obtain a path planning result; The path planning result is converted into a line control chassis execution instruction to control the vehicle to complete the driving behavior; The generative AI model is trained based on a driving planning open source dataset, an extreme scenario dataset, and a traffic flow dataset on a diffusion model using a spatio-temporal attention mechanism. The extreme scenario dataset is obtained based on dynamic scenario enhancement of a generative adversarial network (GAN).

[0007] Further, the acquisition process of the extreme scenario dataset comprises: Based on the obtained normal weather images, a generator in the GAN is used to generate simulated extreme weather images; Discriminate the real extreme weather image and the simulation extreme weather image in the discriminator in the GAN to obtain a discrimination result; Based on the discrimination result, iteratively optimize the generator parameters until the discriminator cannot distinguish the simulation extreme weather image generated by the generator; Add the simulation extreme weather image finally generated by the generator to the extreme scene data set.

[0008] Further, the pre-construction process of the road network quantum graph comprises: The intersection, lane line information and obstacle position in the road network data are taken as a plurality of nodes in the road network quantum graph; Based on the actual road connection between each node, the road section in the road network data is taken as the edge corresponding to the node in the road network quantum graph, and the path selection probability is taken as the edge weight in the road network quantum graph; Based on the speed limit information, the number of lanes and the turning radius in the road network data, the penalty term of the quantum Hamiltonian is converted; Based on the plurality of nodes, the edge corresponding to each node, the edge weight and the penalty term, the road network quantum graph is constructed.

[0009] Further, based on the trajectory prediction result, the obtained map traffic data and the current driving data, the pre-constructed road network quantum graph is solved by using a quantum optimization algorithm to obtain a path planning result, comprising: Based on the trajectory prediction result, the obtained map traffic data and the current driving data, a multi-objective optimization function considering safety, efficiency and energy consumption is used to solve the pre-constructed road network quantum graph by using a quantum optimization algorithm to obtain the path planning result of the vehicle.

[0010] Further, after the trajectory prediction result, the obtained map traffic data and the current driving data are used to solve the pre-constructed road network quantum graph by using a quantum optimization algorithm to obtain a path planning result, further comprising: If the obstacle collision risk event is detected by the terahertz radar, a dynamic re-planning process is triggered, and the re-planning process comprises: Based on the terahertz radar detection data, the trajectory prediction result, the map traffic data and the current driving data, the road network quantum graph is updated; A quantum Monte Carlo tree search algorithm is used to explore a plurality of obstacle-avoiding feasible paths in the updated road network quantum graph in parallel to generate a plurality of alternative path planning results; Based on the safety redundancy constraint and the efficiency loss constraint, the optimal path planning result is screened from the plurality of alternative path planning results.

[0011] Furthermore, the current driving data includes lidar point cloud data, camera images, and millimeter-wave radar data; if an obstacle collision risk event is detected using terahertz radar, it also includes: The echo signal is analyzed from the terahertz radar detection data to obtain the obstacle reflection characteristics; Based on the obstacle reflection characteristics and the preset polarization feature library, the material characteristics of the obstacle are identified using a cosine similarity matching algorithm; The material features and the contour features of the lidar data are fused using low-level features to obtain the low-level fusion result. The semantic features of the camera image and the velocity features of the millimeter radar data are fused using high-level features to obtain a high-level fusion result; Based on the underlying fusion results and the upper-level fusion results, a three-dimensional semantic map including the location, speed, material, and semantic tags of obstacles is generated, and obstacle warnings are issued.

[0012] Furthermore, after obtaining the path planning result by solving the pre-constructed road network quantum graph using a quantum optimization algorithm based on the trajectory prediction result, the acquired map traffic data, and the current driving data, the method further includes: Using an integrated air-space-ground-sea 6G communication network, a cooperative perception message is broadcast to following vehicles in a multi-vehicle cooperative formation. The cooperative perception message includes the path planning result and / or the formation control command based on the path planning result. The process of acquiring the map traffic data also includes: The map traffic data is acquired using an integrated air-space-ground-sea 6G communication network, which includes road map messages and traffic light messages.

[0013] This invention also provides an intelligent driving system based on generative AI and quantum computing, comprising: The trajectory prediction module is used to predict the multimodal behavior trajectory of traffic participants within a preset time period based on the vehicle's current driving data and a generative AI model, thereby obtaining the trajectory prediction result. The path planning module is used to solve the pre-constructed road network quantum graph using a quantum optimization algorithm based on the trajectory prediction results, the acquired map traffic data, and the current driving data, to obtain the path planning results. The driving control module is used to convert the path planning results into drive-by-wire chassis execution commands to control the vehicle to complete driving behavior; The generative AI model is obtained by training a diffusion model with a spatiotemporal attention mechanism on an open-source driving planning dataset, an extreme scenario dataset, and a traffic flow dataset; the extreme scenario dataset is obtained by performing dynamic scene enhancement based on a generative adversarial network (GAN).

[0014] Furthermore, it also includes an acquisition module for: Based on the acquired normal weather images, simulated extreme weather images are generated using the generator in the GAN. The discriminator in the GAN is used to distinguish between real extreme weather images and simulated extreme weather images to obtain the discrimination result; Based on the discrimination result, the generator parameters are iteratively optimized until the discriminator can no longer distinguish the simulated extreme weather images generated by the generator. The simulated extreme weather images finally generated by the generator are added to the extreme scene dataset.

[0015] Furthermore, it also includes building modules for: The intersections, lane line information, and obstacle locations in the road network data are used as multiple nodes in the road network quantum graph; Based on the actual road connection between each node, the road segments in the road network data are used as the edges corresponding to the nodes in the road network quantum graph, and the path selection probability is used as the edge weight in the road network quantum graph. The penalty term is based on the speed limit information, number of lanes, and turning radius in the road network data, which are converted into quantum Hamiltonians. The road network quantum graph is constructed based on the multiple nodes, the edges corresponding to each node, the weights of each edge, and the penalty term.

[0016] Furthermore, the path planning module is specifically used for: Based on the trajectory prediction results, the acquired map traffic data, and the current driving data, and combined with a multi-objective optimization function that considers safety, efficiency, and energy consumption, a quantum optimization algorithm is used to solve the pre-constructed road network quantum network to obtain the vehicle's path planning results.

[0017] Furthermore, the path planning module is also used for: If an obstacle collision risk event is detected using terahertz radar, a dynamic replanning process is triggered, which includes: The road network quantum map is updated based on terahertz radar detection data, the trajectory prediction results, map traffic data, and the current driving data; The quantum Monte Carlo tree search algorithm is adopted, and multiple obstacle avoidance feasible paths are explored in parallel in the updated road network quantum graph using quantum parallelism, generating multiple alternative path planning results; Based on safety redundancy constraints and efficiency loss constraints, the optimal path planning result is selected from the multiple alternative path planning results.

[0018] Furthermore, the current driving data includes lidar point cloud data, camera images, and millimeter-wave radar data; it also includes a warning module for: The echo signal is analyzed from the terahertz radar detection data to obtain the obstacle reflection characteristics; Based on the obstacle reflection characteristics and the preset polarization feature library, the material characteristics of the obstacle are identified using a cosine similarity matching algorithm; The material features and the contour features of the lidar data are fused using low-level features to obtain the low-level fusion result. The semantic features of the camera image and the velocity features of the millimeter radar data are fused using high-level features to obtain a high-level fusion result; Based on the underlying fusion results and the upper-level fusion results, a three-dimensional semantic map including the location, speed, material, and semantic tags of obstacles is generated, and obstacle warnings are issued.

[0019] Furthermore, it also includes a collaboration module for: Using an integrated air-space-ground-sea 6G communication network, a cooperative perception message is broadcast to following vehicles in a multi-vehicle cooperative formation. The cooperative perception message includes the path planning result and / or the formation control command based on the path planning result. The acquisition module is also used for: The map traffic data is acquired using an integrated air-space-ground-sea 6G communication network, which includes road map messages and traffic light messages.

[0020] This application also provides an electronic device, which includes at least a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the steps of the intelligent driving method based on generative AI and quantum computing as described above.

[0021] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the intelligent driving method based on generative AI and quantum computing as described above.

[0022] This application also provides a computer program product, which includes: computer program code, and when the computer program code is run on a computer, causing the computer to perform the steps of any of the above-described intelligent driving methods based on generative AI and quantum computing.

[0023] In this embodiment of the invention, based on the vehicle's current driving data, a generative AI model is used to predict the multimodal behavioral trajectories of traffic participants within a preset future time period, thus obtaining trajectory prediction results. Based on the trajectory prediction results, the acquired map traffic data, and the current driving data, a quantum optimization algorithm is used to solve the pre-constructed road network quantum graph, resulting in path planning results. The path planning results are then converted into drive-by-wire chassis execution commands to control the vehicle to complete driving behavior. The generative AI model is trained on a diffusion model employing a spatiotemporal attention mechanism based on an open-source driving planning dataset, an extreme scenario dataset, and a traffic flow dataset. The extreme scenario dataset is obtained through dynamic scene enhancement using a generative adversarial network (GAN). The generative AI model can predict the multimodal behavioral trajectories of traffic participants in real time and efficiently complete path optimization calculations using quantum computing, thereby improving the accuracy and real-time performance of path planning. Attached Figure Description

[0024] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 A flowchart illustrating an intelligent driving method based on generative AI and quantum computing, provided in an embodiment of the present invention; Figure 2 A flowchart illustrating the structure of a 6G collaborative intelligent driving system based on generative AI and quantum computing, provided in this embodiment of the invention; Figure 3 A schematic diagram of the structure of an intelligent driving system based on generative AI and quantum computing, provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0025] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0026] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0027] Example 1: Figure 1This is a flowchart illustrating an intelligent driving method based on generative AI and quantum computing, provided by an embodiment of the present invention. The process includes the following steps: S101: Based on the vehicle's current driving data, a generative AI model is used to predict the multimodal behavioral trajectories of traffic participants within a preset future time period, and the trajectory prediction results are obtained. S102: Based on the trajectory prediction results, the acquired map traffic data, and the current driving data, the quantum optimization algorithm is used to solve the pre-constructed road network quantum graph to obtain the path planning results; S103: The path planning results are converted into drive-by-wire chassis execution commands to control the vehicle to complete driving behavior; the generative AI model is trained on a diffusion model with spatiotemporal attention mechanism based on open-source driving planning datasets, extreme scenario datasets and traffic flow datasets; the extreme scenario dataset is obtained by dynamic scene enhancement based on Generative Adversarial Network (GAN).

[0028] In this embodiment of the invention, the generative AI model can predict the multimodal behavioral trajectories of traffic participants in real time and efficiently complete path optimization calculations using quantum computing, thus improving the accuracy and real-time performance of path planning.

[0029] The core function of the generative AI model in step 101 above is to predict in real time the multimodal behavioral trajectories of traffic participants (such as pedestrians, vehicles, cyclists, etc.) within a preset time period in the future (e.g., but not limited to the next 5-10 seconds) to support intelligent decision-making in dynamic scenarios.

[0030] The input data package for the generative AI model is current driving data, including but not limited to LiDAR point cloud data, camera images, and millimeter-wave radar data. LiDAR point cloud data includes the environmental contours of the driving environment; camera images include visual semantic information, such as lane lines and traffic signs; and millimeter-wave radar data includes the vehicle's target speed and azimuth angle.

[0031] Optionally, after acquiring the current driving data, preprocessing can be performed on the data. For example, for LiDAR point clouds, the preprocessing process may include: removing rain and snow noise points through median filtering while retaining effective obstacle contours; for camera images, the preprocessing process may include: using the multi-scale Retinex algorithm (an image enhancement algorithm) to enhance image contrast in low-light (nighttime) and strong backlight scenes, thereby improving semantic segmentation accuracy; for millimeter-wave radar, the preprocessing process may include: smoothing velocity data through Kalman filtering to eliminate measurement jitter (error ≤ 0.2 m / s).

[0032] The generative AI model is structured as a diffusion model (DiffusionTransformer) architecture that employs a spatiotemporal attention mechanism to fuse the temporal dimension (e.g., the trajectory of the past 3 seconds) with the spatial dimension (relative positions of multiple targets) into a feature vector.

[0033] The datasets used in training generative AI models include open-source driving planning datasets and traffic flow datasets. Open-source driving planning datasets include, but are not limited to, the nuPlan public dataset (containing urban roads and highway scenarios), while traffic flow datasets include, but are not limited to, 200 hours of actual data collected from logistics vehicles (focusing on complex intersections and construction sections). In addition, to address the scarcity of real-world data for extreme scenarios (heavy rain, smog, nighttime construction), the datasets used in training can also include extreme scenario datasets.

[0034] The loss functions used in the training of generative AI models include mean square error (MSE) and interaction behavior consistency loss function. MSE includes the mean square error between the predicted trajectory and the real trajectory, while interaction behavior consistency loss is used to ensure that there is no conflict between multiple target trajectories.

[0035] The training hardware used in the generative AI model training process includes, but is not limited to, an 8-card NVIDIA A100 GPU cluster in the cloud, a training cycle of 14 days, and 1 million iterations.

[0036] After the generative AI model is trained, the inference process includes: inputting preprocessed multi-sensor data and outputting multimodal trajectories for the next 5 seconds (each modality corresponds to a possible behavior, such as a pedestrian "going straight", "turning left", or "staying"). The inference speed is as follows: 20Hz real-time inference is achieved on an onboard NVIDIA H200 GPU, with a single frame processing time ≤50ms and a prediction error ≤0.3m.

[0037] In this embodiment of the invention, GAN is used for data augmentation to enhance dynamic scenes. GAN includes a generator and a discriminator. The generator takes a normal weather image as input and generates extreme weather images corresponding to the scene (e.g., overlaying raindrops or fog effects onto a sunny image). The discriminator distinguishes between real extreme weather images and generated images, improving the realism of the generated data through adversarial training. In one implementation, the process of acquiring the extreme scene dataset may include: generating simulated extreme weather images using the generator in the GAN based on the acquired normal weather images; using the discriminator in the GAN to distinguish between real extreme weather images and simulated extreme weather images to obtain a discrimination result; iteratively optimizing the generator parameters based on the discrimination result until the discriminator can no longer distinguish the simulated extreme weather images generated by the generator; and adding the finally generated simulated extreme weather images to the extreme scene dataset. Adding the generated extreme scene dataset to the training set can improve the trajectory prediction accuracy of the generative AI model in extreme scenes by 25%.

[0038] In some possible scenarios, the prediction time range of the trajectory prediction results in S101 above can be 5-10 seconds (dynamically adjustable, extended to 10 seconds in high-speed scenarios and shortened to 5 seconds at urban intersections); the number of multimodal outputs can be 3-5 possible trajectories for each traffic participant; it can adapt to extreme scenarios such as heavy rain and nighttime, and the prediction accuracy remains ≥85%. Specifically, the trajectory prediction results include the multimodal trajectory of the traffic participant in the next 5-10 seconds (including location, speed, and acceleration information), and the trajectory confidence score for each trajectory (used as a weight reference for subsequent path optimization).

[0039] The quantum computing path optimization method used in S102 above can realize multi-objective (safety, efficiency, energy consumption) path planning of high-dimensional road networks in milliseconds.

[0040] In one implementation, the pre-construction process of the road network quantum graph in S012 above may include: using intersections, lane line information, and obstacle positions in the road network data as multiple nodes in the road network quantum graph; using road segments in the road network data as edges corresponding to nodes in the road network quantum graph based on the actual road connection between each node, and using path selection probabilities as edge weights in the road network quantum graph; converting speed limit information, number of lanes, and turning radius in the road network data into penalty terms for quantum Hamiltonians; and constructing the road network quantum graph based on multiple nodes, edges corresponding to each node, edge weights, and penalty terms.

[0041] In this implementation, the urban road network can be transformed into a quantum graph using abstraction methods based on road network data. In this graph: Node: represents intersections, lane start / end points, and obstacle locations; Edge: represents road segments, and the amplitude of the quantum state corresponds to the path selection probability; Constraints: speed limits, number of lanes, turning radius, etc. are transformed into penalty terms of quantum Hamiltonians.

[0042] In one implementation, in step S102 above, the path planning result of the vehicle can be obtained by using a quantum optimization algorithm to solve the pre-constructed road network quantum network based on the trajectory prediction result, the acquired map traffic data and the current driving data, combined with a multi-objective optimization function that considers safety, efficiency and energy consumption.

[0043] In this implementation, the input data for the quantum optimization algorithm includes: trajectory prediction results, map traffic data, and current driving data. The map traffic data can be a high-precision map containing real-time traffic congestion coefficients. For example, in this embodiment of the invention, map traffic data can be obtained by calling a publicly available map interface. To ensure the real-time performance of intelligent driving control, map traffic data can also be obtained in real time. Current driving data can also include the vehicle's own status, such as its current location and remaining battery power.

[0044] In the multi-objective optimization function, the optimization objective of the safety term is, for example, to maintain a minimum distance of ≥1.5m from other traffic participants; the optimization objective of the efficiency term can be to minimize the total path travel time (in conjunction with real-time traffic conditions); and the optimization objective of the energy consumption term includes calculating the total energy consumption based on a vehicle energy consumption model (such as the energy consumption coefficient for rapid acceleration / deceleration). In this multi-objective optimization function, the sum of the weights of the safety, efficiency, and energy consumption terms is 1, for example, but not limited to, a weight of 0.4 for the safety term, 0.3 for the efficiency term, and 0.2 for the energy consumption term.

[0045] For example, a quantum optimization algorithm could be a quantum approximate optimization algorithm (QAOA). The QAOA algorithm computes the cost function of all possible paths in parallel using quantum superposition states, quickly converging to the global optimum. For instance, in quantum computing, running the QAOA algorithm on a D-Wave Advantage2 quantum processor (e.g., a 50-qubit system) iteratively optimizes the algorithm through two layers of parameterized quantum circuits, outputting a globally optimal path plan. This implementation has a computation time of ≤200ms (covering a 10 square kilometer urban road network), achieving multi-objective path planning for high-dimensional road networks within milliseconds.

[0046] This quantum computing path optimization can also support dynamic replanning in sudden scenarios. For example, after S103 above, if an obstacle collision risk event is detected by terahertz radar, a dynamic replanning process can be triggered. The replanning process includes: updating the road network quantum map based on terahertz radar detection data, trajectory prediction results, map traffic data, and current driving data; using the quantum Monte Carlo tree search algorithm, and leveraging quantum parallelism, exploring multiple feasible obstacle avoidance paths in parallel in the updated road network quantum map to generate multiple alternative path planning results; and selecting the optimal path planning result from the multiple alternative path planning results based on safety redundancy constraints and efficiency loss constraints.

[0047] Obstacle collision risk events serve as triggering conditions for dynamic replanning. These can include the detection of sudden obstacles by terahertz radar (such as vehicles running red lights or falling objects), and the collision risk with the current path exceeding a preset risk threshold (e.g., >10%). The Quantum Monte Carlo Tree Search (QMCTS) algorithm rapidly expands the search tree through quantum parallelism during the search process, exploring 3-5 alternative paths (avoiding obstacles and complying with traffic rules) (for example only). The technical parameters include: quantum processor, global planning range, dynamic replanning response time, and multi-objective optimization accuracy. The quantum processor can be a D-Wave Advantage2 (e.g., 50 qubits, coherence time ≥100μs). The global planning range can support a maximum road network of 50km × 50km. The dynamic replanning response time can be no more than a set time, for example, ≤50ms (after sudden obstacle detection). Multi-objective optimization accuracy can include a safe distance error ≤0.1m (for example only) and an energy consumption calculation error ≤5% (for example only).

[0048] Safety redundancy constraints are used to limit collision risk within a certain range (e.g., collision risk <1%), while efficiency loss constraints are used to limit the increase in time consumption within a certain range (e.g., collision risk <1%). Based on safety redundancy and efficiency loss, the optimal solution can be selected. For example, the final emergency path can be generated and output within 50ms, and can also be sent synchronously to the cooperative control layer. The optimal path planning result serves as the globally optimal path and can include lane-level navigation instructions: steering angle, target speed, and lane change timing. In addition, emergency alternative paths (including priority ranking) can be output for sudden scenarios, allowing drivers to manually select an emergency path.

[0049] Optionally, the terahertz radar can be replaced with a solid-state blind spot radar (to reduce costs), or an infrared camera can be added to improve nighttime perception capabilities. The path planning algorithm can be replaced with a quantum heuristic algorithm (such as quantum ant colony optimization), which is suitable for scenarios with limited computing power.

[0050] In one possible implementation, a terahertz radar ultra-sensing module can be used to achieve sub-centimeter-level obstacle detection and material recognition in extreme environments such as heavy rain and smog, and to fuse multi-sensor data to generate a high-precision 3D semantic map. For example, the terahertz radar could be a 300GHz photonic chaotic radar deployed in the center of the vehicle roof, with a detection angle of ±120°, a resolution of 0.5cm, and a transmission power of 10mW (compliant with electromagnetic safety standards). It can guarantee a ranging accuracy of ≤0.5cm, a speed range of -100km / h to +200km / h, and the ability to identify ≥10 types of materials (metal, plastic, human body, etc.). In this implementation, if an obstacle collision risk event is detected using terahertz radar, the echo signal of the terahertz radar detection data can be analyzed to obtain the obstacle reflection characteristics. Based on the obstacle reflection characteristics and a preset polarization feature library, the material characteristics of the obstacle are identified using a cosine similarity matching algorithm. The material characteristics and the contour features of the lidar data (such as point cloud density ≥ 200 points / m²) are fused at the lower level to obtain the lower level fusion result. The semantic features of the camera image (such as lane lines and traffic signs) and the velocity features of the millimeter radar data are fused at the higher level to obtain the higher level fusion result. Based on the lower level fusion result and the higher level fusion result, a three-dimensional semantic map including the obstacle's position (x, y, z), velocity (vx, vy, vz), material, and semantic labels is generated, and obstacle warnings are issued.

[0051] When analyzing the echo signals of terahertz radar detection data to obtain obstacle reflection characteristics, the Fourier transform algorithm can be used to extract target distance (based on time-of-flight), velocity (based on the Doppler effect), and polarization characteristics (changes in the electric field direction of the reflected wave) from the terahertz radar echo signal. This extracted information can be used as obstacle reflection characteristics. A pre-set polarization feature library can include reflection feature templates for various materials such as metal, plastic, human body, and water. This allows for the identification of obstacle material characteristics and properties through cosine similarity matching (accuracy > 98%). In extreme environments such as heavy rain, rain clutter suppression algorithms (e.g., wavelet transform-based denoising) can be used to improve the signal-to-noise ratio (SNR ≥ 20dB) and achieve signal compensation in extreme environments.

[0052] In this implementation, various data can be synchronized in time and calibrated spatially before feature fusion. During time synchronization, calibration can be performed using GPS (Global Positioning System) time synchronization combined with the vehicle's clock to ensure that the timestamp error of the terahertz radar, lidar, camera, and millimeter-wave radar is ≤5ms. Spatial calibration includes extrinsic parameter calibration and coordinate unification. In the extrinsic parameter calibration method, a checkerboard calibration board can be used to determine the transformation matrix between each sensor and the vehicle's coordinate system (rotation angle error ≤0.1°, translation error ≤1cm). During coordinate unification, all sensor data can be converted to the vehicle's coordinate system (origin at the front wheel axle). The map accuracy after multi-sensor fusion is: absolute positioning error ≤5cm, obstacle miss rate <0.1%.

[0053] Preferably, in this implementation, performance enhancement can be achieved through the following improvements in extreme environments such as heavy rain and nighttime scenarios. In heavy rain scenarios, the strong penetrating power of terahertz waves (rain and fog attenuation rate <2dB / km), combined with multi-frame data accumulation (3-frame sliding window), can increase the detection distance to >150m. In nighttime scenarios without illumination, the detection distance is >180m. In nighttime scenarios, no active supplemental lighting is required (terahertz waves are unaffected by light). By fusing with an infrared camera (optional), the pedestrian recognition rate can be improved to 99%. Table 1 shows a comparison of the performance of traditional obstacle recognition schemes with the fusion of terahertz radar and lidar proposed in this invention.

[0054] Table 1

[0055] The 3D semantic map in this implementation (e.g., updated 10 times per second, covering a radius of 200m) can enable high-priority obstacle alerts (such as pedestrians suddenly entering or falling metal obstacles).

[0056] In one optional implementation, this embodiment of the invention further includes a 6G global collaborative communication module to construct an integrated air-ground-sea communication network, enabling low-latency data interaction between vehicles (V2V), vehicles (V2I), and vehicles (V2C), supporting multi-vehicle collaborative decision-making. When building the communication network architecture, at the ground layer: 6G integrated sensing and computing base stations are deployed on the roadside, operating in the 300GHz band, with a coverage radius of 2km, supporting simultaneous access for 200 vehicles (each vehicle's bandwidth ≥1Gbps); at the air-space layer: in remote areas (such as highways and rural roads), blind spots are filled using low-orbit satellites (such as Starlink) + ground relay base stations, achieving full coverage without dead zones, with a positioning accuracy ≤0.1m.

[0057] The vehicle-road cooperative data interaction protocol adopts the 3GPP Release 18 standard, defining three types of core messages: Cooperative Awareness Message (CAM): Vehicles periodically (10Hz) broadcast their own position, speed, and trajectory prediction results (for reference by surrounding vehicles); Road Map Message (MAP): Roadside units broadcast real-time traffic conditions (construction sections, temporary lane changes), with an update frequency of 1Hz; Traffic Light Message (SPAT): Traffic light status (red / green / yellow light and remaining time), with an update frequency of 10Hz.

[0058] The platooning logic in multi-vehicle collaborative platooning control is as follows: Based on 6G real-time communication (latency <20ms), a 100-vehicle platoon maintains a spacing of ≤2m and a speed synchronization error of ≤1km / h. Data interaction process: The lead vehicle (navigator) calculates the global path and broadcasts it to following vehicles; following vehicles fine-tune their speed and steering based on the lead vehicle's trajectory and their own sensor data to ensure platoon stability; in emergency scenarios (such as the lead vehicle braking suddenly): the lead vehicle sends an emergency braking command, and following vehicles respond within 10ms to avoid rear-end collisions. Based on this, after S102, a collaborative perception message can be broadcast to following vehicles in the multi-vehicle collaborative platoon using an integrated air-space-ground-sea 6G communication network. This collaborative perception message includes path planning results and / or platooning control commands based on the path planning results. During the acquisition of map traffic data, the integrated air-space-ground-sea 6G communication network is also used to acquire map traffic data, which includes road map messages and traffic light messages.

[0059] The 6G global collaborative communication module includes: 6G communication rate: peak rate of vehicle terminal >100Gbps, average rate ≥10Gbps; end-to-end latency: vehicle-to-infrastructure communication ≤12ms, vehicle-to-vehicle communication ≤8ms; formation control accuracy: spacing error ±0.5m, vehicle speed synchronization error ≤0.5km / h; coverage capability: 100% coverage in urban areas, ≥99.9% coverage in remote areas. Specifically, it can output at least one of the following data: 3D semantic map (fusion of multi-sensor features), vehicle-to-infrastructure / vehicle-to-vehicle collaborative messages (CAM, MAP, SPAT), and formation control commands (following vehicle speed, steering angle).

[0060] In S103 above, the execution control and safety redundancy module can be used to convert path planning commands into vehicle execution actions and trigger human-machine collaboration in high-risk scenarios to ensure driving safety. Specifically, the following drive-by-wire chassis execution control can be executed: Steering control: Receives the steering angle command from the path planning and achieves steering accuracy of ±0.1° through the drive-by-wire steering system (EPS), with a response delay ≤30ms; Braking / acceleration control: Calculates braking pressure (accuracy ±0.5bar) or acceleration torque based on the target vehicle speed to ensure that the vehicle speed control error is ≤1km / h.

[0061] This invention can also achieve risk level assessment and human-machine collaboration. For example, risk level classification (levels 1-10): based on a comprehensive assessment of obstacle distance, relative speed, and path conflict probability (e.g., when the distance is <5m and the relative speed is >30km / h, the risk level is 9). High-risk scenario response (level >8): multimodal warning: red flashing HUD (head-up display) (visual), buzzer alarm (auditory, frequency 2kHz), seat vibration (tactile, intensity 50Hz); manual takeover window: a 300ms window is reserved for driver takeover; if there is no response, the system automatically performs emergency braking (deceleration ≥8m / s²).

[0062] In this implementation process, the steering control accuracy is ±0.1°, with a response delay of ≤30ms; braking pressure accuracy is ±0.5bar, with an emergency braking distance (100km / h→0) ≤40m; risk assessment accuracy is ≥99.5% (no missed alarms, false alarm rate <0.1%); and manual intervention response time is ≤300ms (system reserved window). Output of drive-by-wire chassis execution commands (steering angle, braking pressure, acceleration torque); and human-machine interaction warning signals (visual / auditory / tactile).

[0063] Embodiments of the present invention may further include a cloud-based training and simulation verification module, which optimizes the model through data closed-loop iteration and verifies the robustness of the system in a digital twin environment, supporting the continuous evolution of the system.

[0064] In the closed-loop data training process, data acquisition involves real vehicles uploading 10TB of anonymized data weekly via 6GB (raw data from multiple sensors, trajectory prediction errors, and path planning results). Data processing includes cleaning (denoising and deduplication) and labeling (manual and automated labeling tools) on a cloud server to build an incremental training set. Model iteration involves retraining the generative AI model and quantum optimization algorithm monthly based on 1 million kilometers of newly added real-world road condition data, improving the model's generalization ability by 15% every quarter.

[0065] In digital twin simulation verification, the virtual environment is constructed using CARLA (an open-source autonomous driving simulation platform) and PyBullet (a physics engine), simulating multiple scenarios including urban areas, highways, and rural areas, supporting the reproduction of extreme weather (typhoons, heavy rain) and sudden events (earthquakes, multi-vehicle collisions). Verification metrics include: system failure rate (≤0.1 times / 1000 km), decision rationality score (≥9.0 / 10 points), and extreme scenario survival rate (≥99.9%). Model deployment: The model verified through simulation is pushed to the vehicle terminal via OTA (Over-The-Air) updates (minor updates every two weeks, major updates every month).

[0066] In the cloud-based training and simulation verification module, the data acquisition volume is as follows: 1 million kilometers of real-world road condition data are added monthly, with a storage capacity of 1PB / year; the model training cycle is 7 days / cycle for generative AI models and 14 days / cycle for quantum optimization algorithms; the simulation scenario coverage includes ≥1000 typical scenarios and 50 extreme scenarios; and the OTA update success rate is ≥99.9% (with breakpoint resumption and verification mechanisms). Output data includes: iteratively optimized generative AI model and quantum algorithm parameters; simulation verification report (system performance indicators); and OTA update package (for download by the vehicle terminal).

[0067] Through the collaborative work of the above modules, the system realizes a closed loop of the entire process of "environmental perception - dynamic modeling - path optimization - collaborative control - safe execution - model evolution", achieving dual optimization of safety and efficiency in complex road environments.

[0068] Figure 2 This paper presents an intelligent driving system that integrates generative AI dynamic modeling, quantum computing path optimization, terahertz radar ultra-sensing, and 6G global collaborative communication, suitable for real-time decision-making and multi-vehicle collaboration in complex road environments.

[0069] The system hardware adopts an integrated design of "multi-sensor fusion + quantum-classical heterogeneous computing + 6G omni-channel communication" to ensure high efficiency in environmental perception, data processing, and collaborative interaction. This includes sensor deployment and communication modules.

[0070] Sensor deployment meets the following requirements: Terahertz radar: Installed in the center of the vehicle roof, it uses a 300GHz photonic chaotic radar. Key parameters include: detection angle ±120° (full horizontal coverage), resolution 0.5cm (sub-centimeter level obstacle recognition), and transmission power 10mW (compliant with electromagnetic safety standards). Its core function is to penetrate obstructions in extreme environments such as heavy rain and fog, while simultaneously extracting the material characteristics of obstacles (such as metal, human body, plastic, etc.).

[0071] LiDAR: The front grille integrates a 128-line LiDAR (main radar) with a ranging range of 200m and an accuracy of ±2cm, used to capture environmental contours and static obstacles (such as guardrails and curbs); four blind spot LiDARs are deployed on the sides of the vehicle (two on each side) to cover 180° blind spots on both sides of the vehicle, avoiding the risk of side collisions.

[0072] Other auxiliary sensors: Cameras: An 8-megapixel wide-angle camera (120° field of view) is deployed above the windshield, and four high-definition cameras are deployed on the side windows to recognize visual semantic information such as lane lines and traffic signs; Millimeter-wave radar: Two 77GHz millimeter-wave radars are embedded in each of the front and rear bumpers, focusing on capturing the speed (error ≤0.2m / s) and azimuth of dynamic targets, forming redundancy with the terahertz radar.

[0073] Computing Unit: The vehicle domain controller adopts a "quantum-classical heterogeneous computing platform," with the following core configuration: Quantum processor: D-Wave Advantage2 (50 qubits, coherence time ≥100μs), responsible for quantum computing for high-dimensional path optimization; Classic GPU: NVIDIA H200 (computing power ≥1000TOPS), supporting generative AI model inference and multi-sensor data fusion; the two are interconnected through a high-speed PCIe 5.0 interface, with data interaction latency ≤1ms, meeting real-time processing requirements.

[0074] The communication module achieves seamless data interaction across the entire network by constructing a multi-layered communication network encompassing vehicles, roads, the cloud, and airspace. 6G vehicle terminal: integrated into the vehicle communication box, supporting dual-band communication: (1) Terrestrial communication: operating in the 300GHz terahertz band, with a peak transmission rate of 100Gbps and an end-to-end delay of <10ms (vehicle-to-vehicle, vehicle-to-road interaction); (2) Sky communication: using low-orbit satellites (such as Starlink) to fill blind spots in remote areas, with a positioning accuracy of ≤0.1m, ensuring continuous coverage in scenarios such as highways and rural roads. (3) Roadside unit: deploying one 6G integrated base station every 2km along the road; (4) Communication coverage: supporting simultaneous access of 200 vehicles, with each vehicle allocated a bandwidth of ≥1Gbps to meet the real-time data sharing needs of multiple vehicles.

[0075] Edge computing: Built-in AI acceleration chip can preprocess data from roadside cameras and radar, reducing the computing power consumption of on-board units; Dynamic updates: Real-time broadcast of road conditions (such as construction sections and temporary lane changes) and traffic light status to vehicles, with an update frequency of 1-10Hz (dynamically adjusted according to the scenario).

[0076] The software algorithm is designed around the entire process of "dynamic modeling - quantum optimization - cooperative control", and lightweight deployment is achieved through model iteration and hardware adaptation.

[0077] 2.1 Generative AI Dynamic Modeling The core idea is to achieve multimodal trajectory prediction of dynamic traffic participants by combining a diffusion model with a GAN network. The specific process is as follows: Data preprocessing: (1) Terahertz radar point cloud: Median filtering algorithm is used to remove rain and snow noise points (effective obstacle outlines are preserved, noise removal rate > 95%); (2) Camera images: For low light or strong backlight scenes at night, multi-scale Retinex algorithm is used to enhance contrast (image signal-to-noise ratio is improved by ≥ 20dB) to ensure that the semantic segmentation accuracy of lane lines and traffic signs is > 98%; (3) Millimeter-wave radar data: Kalman filtering is used to smooth the speed measurement value and control the speed jitter error to ≤ 0.2m / s.

[0078] Diffusion model training and inference: (1) Training data: The nuPlan public dataset (urban roads, highway scenarios) and 200 hours of actual data collected by logistics vehicles (focusing on long-tail scenarios such as complex intersections and construction sections) are integrated; (2) Model structure: The Diffusion Transformer architecture with spatiotemporal attention mechanism is adopted. The input is multi-sensor data (time dimension) and relative position of multiple targets (spatial dimension) in the past 3 seconds, and the output is feature vector of the next 5-10 seconds; (3) Training configuration: Cloud 8-card NVIDIA A100 GPU cluster, training cycle of 14 days, 1 million iterations, and the loss function adopts "mean square error (MSE) between predicted trajectory and real trajectory + interaction behavior consistency loss" (to ensure that there is no conflict in the trajectory of multiple targets); (4) Inference performance: 20Hz real-time inference is achieved on the vehicle NVIDIA H200 GPU (single frame processing time ≤50ms), prediction error ≤0.3m, and supports 3-5 possible trajectories for each traffic participant (such as pedestrian "straight", "left turn" and "stop").

[0079] Extreme Scene Enhancement (GAN Network): To address the scarcity of extreme weather (heavy rain, fog and haze) data, extreme scene data is synthesized through Generative Adversarial Network (GAN): (1) Generator: Input normal weather images and overlay special effects such as raindrops and fog to generate realistic extreme scene images; (2) Discriminator: Distinguish between real and generated images, and improve the authenticity of generated data through 100,000 adversarial training sessions; (3) Effect: After adding the generated data to the training set, the trajectory prediction accuracy of the model in extreme scenes is improved by 25% (from 60% to 85%).

[0080] 2.2 Quantum computing path optimization: Real-time path planning and dynamic replanning of high-dimensional road networks are achieved through quantum graph modeling and quantum algorithms. Quantum modeling of road networks: Abstracting the urban road network into a "quantum graph", where: nodes: represent intersections, lane start / end points, and obstacle locations; Edge: Represents a road segment; the amplitude of the quantum state corresponds to the path selection probability; Constraints: convert speed limits, number of lanes, turning radius, etc. into penalty terms of the quantum Hamiltonian (e.g., penalty weight + 0.5 for speeding scenarios).

[0081] Global Path Planning (QAOA Algorithm): Input: Trajectory prediction results from generative AI, high-precision map (including real-time traffic congestion coefficient), vehicle status (current location, remaining battery power); Objective function: Safety term (weight 0.4, minimum distance from other traffic participants ≥ 1.5m), efficiency term (weight 0.3, shortest total path time), energy consumption term (weight 0.2, calculated based on energy consumption coefficients during rapid acceleration / deceleration); Calculation process: Two layers of parameterized quantum circuits are run on the D-Wave Advantage2 quantum processor (50 qubits), and convergence to the global optimum is achieved through iterative optimization, with a time ≤ 180ms (covering a 10 square kilometer urban road network).

[0082] Dynamic Replanning (QMCTS Algorithm): Triggered when the terahertz radar detects a sudden obstacle (such as a vehicle running a red light) and the collision risk is >10%; Quantum Monte Carlo Tree Search: The search tree is rapidly expanded through quantum parallelism, exploring 3-5 alternative paths within 50ms; Filtering criteria: safety redundancy (collision risk <1%), efficiency loss (time increase ≤10% compared to the original path), and finally outputting 1 optimal emergency path.

[0083] 2.3 6G Global Collaborative Communication: Based on the 6G network, low-latency data interaction between vehicles, between vehicles and roads, and between vehicles and the cloud is achieved, supporting multi-vehicle collaborative decision-making. Communication Protocol: Adopting the 3GPP Release 18 standard, three types of core messages are defined: CAM (Collaborative Awareness Message): Every 100ms (10Hz), vehicles broadcast their own position, speed, and trajectory prediction results for reference by surrounding vehicles; MAP (Road Map Message): Roadside units broadcast real-time road conditions (such as construction sections and temporary lane changes) every 1 second (1Hz); SPAT (Traffic Light Message): Traffic light status (red / green / yellow light and remaining time) is updated every 100ms (10Hz) to ensure that vehicles adjust their speed in advance.

[0084] Multi-vehicle collaborative platooning control: Platooning logic: The lead vehicle calculates the global path and broadcasts it to the following vehicles. Through 6G communication (delay < 20ms), the spacing between 100 vehicles in the platoon is maintained at ≤ 2m, and the vehicle speed synchronization error is ≤ 1km / h. Emergency scenario response: When the lead vehicle brakes suddenly, the emergency braking command is transmitted to all following vehicles through the 6G network within 10ms to achieve synchronized braking (deceleration ≥ 8m / s²) and avoid rear-end collisions.

[0085] 3. Test and verification data: Through real vehicle testing and simulation verification, the core performance indicators of the system are as follows: Dynamic prediction accuracy: On the nuScenes dataset, the prediction error of pedestrian trajectory in the next 5 seconds is ≤0.28m, and the decision rationality score of multi-vehicle interaction scenarios (such as meeting at intersections) reaches 9.2 / 10 (expert evaluation). Path planning efficiency: The quantum QAOA algorithm takes 180ms to plan routes in complex urban road networks (including 1000+ intersections), which is more than 100 times faster than the traditional A* algorithm (20 seconds); Extreme environment performance: Terahertz radar has a detection range >150m and a material recognition accuracy >98% in heavy rain scenarios; a detection range >180m in low-light nighttime scenarios, with a blind spot coverage rate <0.8%; Cooperative communication latency: Vehicle-to-road communication latency ≤12ms, vehicle-to-vehicle communication latency ≤8ms, supporting platooning of 100 vehicles with spacing control accuracy ±0.5m. Through the above hardware deployment and algorithm implementation, the system can achieve a closed-loop process of "perception-prediction-planning-control" in complex road environments, meeting the dual requirements of safety and efficiency.

[0086] The following comparison between traditional solutions and embodiments of the present invention demonstrates the effectiveness of the present invention. Currently, the following traditional solutions exist: Static Environmental Modeling: Traditional deep learning models cannot capture dynamic traffic intentions in real time, leading to lag in path planning. Traditional models cannot predict the multimodal behaviors of traffic participants (e.g., pedestrians may go straight, turn left, or stay), resulting in limited obstacle avoidance strategies and weak adaptability to dynamic scenarios. In this embodiment of the invention, generative AI dynamic modeling predicts the multimodal behavioral trajectories of traffic participants in real time for the next 5-10 seconds. The diffusion model achieves a 92% accuracy rate in predicting pedestrian trajectories, improves the rationality of decision-making in multi-vehicle interaction scenarios by 40%, and enhances the accuracy of dynamic scenario prediction.

[0087] Inefficiency of path optimization: Dijkstra's algorithm has high computational complexity in high-dimensional road networks, making it difficult to meet millisecond-level real-time requirements. End-to-end models require tens of TOPS of computing power, consuming significant computing resources and making lightweight deployment in vehicle domain controllers difficult. In this embodiment of the invention, quantum computing path optimization completes multi-objective (safety, efficiency, energy consumption) path planning for high-dimensional road networks within milliseconds. The quantum annealing algorithm reduces path planning time from 20 seconds in traditional algorithms to 180ms, improves congestion avoidance rate by 35%, and ensures efficient and safe path planning.

[0088] Limitations in perception capabilities: LiDAR performance drops by more than 30% in rain and fog, and millimeter-wave radar cannot identify stationary obstacles. The perception accuracy of a single sensor decreases by more than 30% in scenarios such as heavy rain and fog, posing safety hazards and the risk of failure in extreme environments. In this embodiment of the invention, terahertz radar provides superior perception: achieving sub-centimeter-level obstacle detection and material identification under extreme weather conditions. In heavy rain, the terahertz radar has a detection range of >150m, a material identification accuracy of >98%, and a blind zone coverage of <1%, ensuring reliable perception in extreme environments.

[0089] Collaborative communication latency: Existing vehicle-to-everything (V2X) communication latency is >200ms, which cannot support real-time multi-vehicle platooning. Vehicle-road cooperative information interaction latency is >200ms, which cannot support dynamic multi-vehicle platooning and emergency event linkage, resulting in delayed collaborative decision-making. In this embodiment of the invention, 6G full-domain collaborative communication: an integrated air-space-ground-sea communication network is constructed to support multi-vehicle collaboration and real-time data interaction with the cloud. 6G communication achieves multi-vehicle collaborative decision-making latency <20ms, supports dynamic platooning of 100 vehicles, improves traffic efficiency by 50%, and enables real-time linkage across the entire domain; the cloud training platform adds 1 million kilometers of real road condition data every month, and the model generalization ability improves by 15% every quarter, ensuring the system's continuous evolution capability.

[0090] In summary, traditional technologies cannot simultaneously address dynamic scene prediction, high real-time path planning, extreme environment perception, and low-latency collaborative communication, resulting in safety hazards and efficiency bottlenecks in intelligent driving systems in complex scenarios. This invention aims to solve the integrated problem of dynamic environment modeling, high-dimensional path optimization, extreme perception, and low-latency collaboration to achieve safe and efficient intelligent driving.

[0091] Example 2: Based on the same concept, Figure 2 A schematic diagram of a smart driving system based on generative AI and quantum computing, provided as an embodiment of the present invention, includes: The trajectory prediction module is used to predict the multimodal behavior trajectory of traffic participants within a preset time period based on the vehicle's current driving data and a generative AI model, thereby obtaining the trajectory prediction result. The path planning module is used to solve the road network quantum graph using a pre-built quantum optimization algorithm based on the trajectory prediction results, the acquired map traffic data, and the current driving data, to obtain the path planning results. The driving control module is used to convert the path planning results into drive-by-wire chassis execution commands to control the vehicle to complete driving behavior; The generative AI model is obtained by training a diffusion model with a spatiotemporal attention mechanism on an open-source driving planning dataset, an extreme scenario dataset, and a traffic flow dataset; the extreme scenario dataset is obtained by performing dynamic scene enhancement based on a generative adversarial network (GAN).

[0092] In one possible implementation, an acquisition module is also included, for: Based on the acquired normal weather images, simulated extreme weather images are generated using the generator in the GAN. The discriminator in the GAN is used to distinguish between real extreme weather images and simulated extreme weather images to obtain the discrimination result; Based on the discrimination result, the generator parameters are iteratively optimized until the discriminator can no longer distinguish the simulated extreme weather images generated by the generator. The simulated extreme weather images finally generated by the generator are added to the extreme scene dataset.

[0093] In one possible implementation, a building module is also included for: The intersections, lane line information, and obstacle locations in the road network data are used as multiple nodes in the road network quantum graph; Based on the actual road connection between each node, the road segments in the road network data are used as the edges corresponding to the nodes in the road network quantum graph, and the path selection probability is used as the edge weight in the road network quantum graph. The penalty term is based on the speed limit information, number of lanes, and turning radius in the road network data, which are converted into quantum Hamiltonians. The road network quantum graph is constructed based on the multiple nodes, the edges corresponding to each node, the weights of each edge, and the penalty term.

[0094] In one possible implementation, the path planning module is specifically used for: Based on the trajectory prediction results, the acquired map traffic data, and the current driving data, and combined with a multi-objective optimization function that considers safety, efficiency, and energy consumption, a quantum optimization algorithm is used to solve the pre-constructed road network quantum network to obtain the vehicle's path planning results.

[0095] In one possible implementation, the path planning module is further configured to: If an obstacle collision risk event is detected using terahertz radar, a dynamic replanning process is triggered, which includes: The road network quantum map is updated based on terahertz radar detection data, the trajectory prediction results, map traffic data, and the current driving data; The quantum Monte Carlo tree search algorithm is adopted, and multiple obstacle avoidance feasible paths are explored in parallel in the updated road network quantum graph using quantum parallelism, generating multiple alternative path planning results; Based on safety redundancy constraints and efficiency loss constraints, the optimal path planning result is selected from the multiple alternative path planning results.

[0096] In one possible implementation, the current driving data includes lidar point cloud data, camera images, and millimeter-wave radar data; it also includes a warning module for: The echo signal is analyzed from the terahertz radar detection data to obtain the obstacle reflection characteristics; Based on the obstacle reflection characteristics and the preset polarization feature library, the material characteristics of the obstacle are identified using a cosine similarity matching algorithm; The material features and the contour features of the lidar data are fused using low-level features to obtain the low-level fusion result. The semantic features of the camera image and the velocity features of the millimeter radar data are fused using high-level features to obtain a high-level fusion result; Based on the underlying fusion results and the upper-level fusion results, a three-dimensional semantic map including the location, speed, material, and semantic tags of obstacles is generated, and obstacle warnings are issued.

[0097] In one possible implementation, a collaboration module is also included, for: Using an integrated air-space-ground-sea 6G communication network, a cooperative perception message is broadcast to following vehicles in a multi-vehicle cooperative formation. The cooperative perception message includes the path planning result and / or the formation control command based on the path planning result. The acquisition module is also used for: The map traffic data is acquired using an integrated air-space-ground-sea 6G communication network, which includes road map messages and traffic light messages.

[0098] Example 3: Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Based on the above embodiments, this embodiment of the present invention also provides an electronic device, including a processor 401, a communication interface 402, a memory 403, and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404. The memory 403 stores a computer program. When the program is executed by the processor 401, the processor 401 performs the intelligent driving method based on generative AI and quantum computing as described above.

[0099] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0100] Communication interface 402 is used for communication between the above-mentioned electronic device and other devices.

[0101] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0102] The processors mentioned above can be general-purpose processors, including central processing units, network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0103] Example 4: Based on the above embodiments, this invention also provides a computer-readable storage medium storing a computer program, which is processed by the above-described intelligent driving method based on generative AI and quantum computing.

[0104] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative in all respects and are not the only ones. All modifications within the scope of this invention or its equivalents are included in this invention.

[0105] The embodiments described in this invention are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses.

[0106] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0107] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0108] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0109] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for intelligent driving based on generative AI and quantum computing, characterized in that, The method comprises the following steps: Based on the current driving data of the vehicle, a generative AI model is used to predict the multi-modal behavior trajectory of the traffic participants in a future preset period, and a trajectory prediction result is obtained; Based on the trajectory prediction result, the obtained map traffic data and the current driving data, a quantum optimization algorithm is used to solve a pre-constructed road network quantum graph, and a path planning result is obtained; The path planning result is converted into a drive-by-wire chassis execution instruction to control the vehicle to complete the driving behavior; The generative AI model is trained based on a driving planning open source dataset, an extreme scenario dataset and a traffic flow dataset to obtain a diffusion model with a space-time attention mechanism; The extreme scenario dataset is obtained based on a generative adversarial network (GAN) for dynamic scenario enhancement.

2. The method of claim 1, wherein, The process of obtaining the extreme scenario dataset comprises the following steps: Based on the obtained normal weather images, a generator in the GAN is used to generate simulated extreme weather images; A discriminator in the GAN is used to distinguish between real extreme weather images and the simulated extreme weather images, and a discrimination result is obtained; Based on the discrimination result, the generator parameters are iteratively optimized until the discriminator cannot distinguish between the simulated extreme weather images generated by the generator; The simulated extreme weather images finally generated by the generator are added to the extreme scenario dataset.

3. The method of claim 1, wherein, The pre-construction process of the road network quantum graph comprises the following steps: The intersection, lane line information and obstacle position in the road network data are taken as a plurality of nodes in the road network quantum graph; Based on the actual road connection between each node, the road segments in the road network data are taken as the edges corresponding to the nodes in the road network quantum graph, and the path selection probability is taken as the edge weight in the road network quantum graph; Based on the speed limit information, the number of lanes and the turning radius in the road network data, a penalty term of the quantum Hamiltonian is converted; Based on the plurality of nodes, the edges corresponding to each node, the edge weights and the penalty term, the road network quantum graph is constructed.

4. The method of claim 1 or 3, wherein, Based on the trajectory prediction result, the obtained map traffic data and the current driving data, a quantum optimization algorithm is used to solve the pre-constructed road network quantum graph, and a path planning result is obtained, which comprises the following steps: Based on the trajectory prediction result, the obtained map traffic data and the current driving data, a multi-objective optimization function considering safety, efficiency and energy consumption is used to solve the pre-constructed road network quantum graph by a quantum optimization algorithm, and a path planning result of the vehicle is obtained.

5. The method of claim 1, wherein, After the trajectory prediction result, the obtained map traffic data and the current driving data are used to solve the pre-constructed road network quantum graph by a quantum optimization algorithm, a path planning result is obtained, and the method further comprises the following steps: If an obstacle collision risk event is detected by a terahertz radar, a dynamic re-planning process is triggered, which comprises the following steps: Based on the terahertz radar detection data, the trajectory prediction result, the map traffic data and the current driving data, the road network quantum graph is updated; A quantum Monte Carlo tree search algorithm is used to explore a plurality of obstacle-avoiding feasible paths in parallel in the updated road network quantum graph by quantum parallelism, and a plurality of alternative path planning results are generated; Based on the safety redundancy constraint and the efficiency loss constraint, the optimal path planning result is screened from the multiple alternative path planning results.

6. The method of claim 1 or 5, wherein, The current driving data includes lidar point cloud data, camera image and millimeter wave radar data; if a terahertz radar is used to detect an obstacle collision risk event, the current driving data further includes: Performing echo signal analysis on the terahertz radar detection data to obtain obstacle reflection characteristics; Based on the obstacle reflection characteristics and a preset polarization feature library, a cosine similarity matching algorithm is used to identify the material characteristics of the obstacle; Performing bottom layer feature fusion on the material characteristics and the contour characteristics of the lidar data to obtain a bottom layer fusion result; Performing high layer feature fusion on the semantic characteristics of the camera image and the speed characteristics of the millimeter wave radar data to obtain a high layer fusion result; Based on the bottom layer fusion result and the high layer fusion result, a three-dimensional semantic map including the position, speed, material and semantic label of the obstacle is generated, and an obstacle warning is performed.

7. The method of claim 1, wherein, After the path planning result is obtained by solving the pre-constructed road network quantum graph based on the trajectory prediction result, the obtained map traffic data and the current driving data, the method further includes: Using a space-air-ground-sea integrated 6G communication network, a cooperative perception message including the path planning result and / or a formation control instruction of the path planning result is broadcast to a following vehicle in a multi-vehicle cooperative formation; The process of obtaining the map traffic data further includes: Using a space-air-ground-sea integrated 6G communication network, the map traffic data including road map messages and signal lamp messages is obtained.

8. An intelligent driving system based on generative AI and quantum computing, characterized in that, The method includes: a trajectory prediction module configured to predict a multi-modal behavior trajectory of a traffic participant in a future preset time period based on current driving data of a vehicle using a generative AI model to obtain a trajectory prediction result; a path planning module configured to solve a pre-constructed road network quantum graph using a quantum optimization algorithm based on the trajectory prediction result, obtained map traffic data and the current driving data to obtain a path planning result; a driving control module configured to convert the path planning result into a drive-by-wire chassis execution instruction to control the vehicle to complete a driving behavior; The generative AI model is trained based on a driving planning open source dataset, an extreme scenario dataset and a traffic flow dataset on a diffusion model using a spatio-temporal attention mechanism; The extreme scenario dataset is obtained based on a generative adversarial network (GAN) for dynamic scenario enhancement.

9. An electronic device, comprising: The electronic device includes a processor and a memory, the memory is used to store program instructions, and the processor is used to execute the computer program stored in the memory to implement the steps of any one of the above methods based on generative AI and quantum computing.

10. A computer-readable storage medium, characterized in that, The computer program is stored in the memory and executed by the processor to implement the steps of any one of the above methods based on generative AI and quantum computing.

Images